Weighting images in dataloader

[MLRadfys]

Hi,

I have a dataset which I want to use for object detection. The dataset is a combination of one larger dataset and two smaller ones.
Now I want to weight the images on a batch-level, so that each batch contains 50% more data from the larger subset.

So for example:
batch_size = 12

one batch should contain 6 images from the large subset and 3 from each of the smaller ones.

Is that something that can be done with Detectron2 ?

I was thinking that maybe the RepeatFactorTrainingSampler could be a solution, by e.g. setting the repeat_factor equal to 2 for images contained in the larger subset and to 1 for the other subsets.

Thanks in advance for any suggestion,

kind regards,

M